Predicting network performance with high granularity is essential for optimizing resource utilization and ensuring Quality of Service (QoS) in modern communication networks. Traditional models struggle to capture the fine-grained dynamics of packet arrivals, particularly under conditions of high traffic and network variability. This project proposes a transformer-based predictive framework that leverages temporal point processes to model network events with extreme precision, focusing on both when and how packets will arrive. Transformers have demonstrated exceptional capabilities in capturing long-range dependencies and complex temporal patterns, making them well-suited for the high-dimensional and time-dependent nature of network events. Given the size of network data and the need for high-resolution predictions, GPU acceleration will be a critical component of the project. The transformer architecture is computationally intensive, particularly when dealing with the long sequences typical in network performance data. GPUs will enable us to train and fine-tune the model efficiently, significantly reducing the time required to process and learn from large-scale network datasets. By leveraging GPUs, we can handle larger sequences, increase model complexity, and perform more granular predictions without sacrificing performance. The expected outcomes include a transformer-based model that can predict packet arrivals and network performance with extreme precision, enabling better traffic management, enhanced QoS, and proactive detection of network bottlenecks. The model will be validated through simulations and real-world network data, demonstrating its ability to deliver actionable insights and improve network resilience.